The　paper　discusses　an　application　of　multi-agent　based　on　theory　of　argumentation-based　on　negotiation,　and　provides　adaptive　concession　strategy　model　for　beginning　a　negotiation.　Firstly,　this　paper　defines　hypotheses　of　model　and　a　frame　of　negotiation　based　on　argumentation.　Secondly,　for　comparing　purpose,　two　generating　models　of　concession　strategy　are　also　studied　as:　model　based　on　time　constraint　and　model　based　on　opposite＇s　preferences,　the　process　of　demonstration　and　result　of　experiment　has　been　shown　that　the　latter　designed　by　PSO-RBFNN　(RBF　Neural　Network　optimized　by　Particle　Swarm　Optimization)　algorithm　has　better　abilities　of　learning　and　reasoning,　which　is　dominant　strategy　in　bilateral　negotiations,　and　has　a　certain　feasibility　and　application　value.